Model training method, transformer fault diagnosis method and related device
By employing a multi-filter semi-supervised feature selection method in transformer fault diagnosis, and using the confidence criterion to select unlabeled DGA data, the training dataset is expanded, which solves the problem of single filter relying on labeled data and improves the accuracy and robustness of fault diagnosis.
Patent Information
- Application Number
- CN202211224772.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-09
- Publication Date
- 2026-02-06
- Estimated Expiration
- 2042-10-09
AI Technical Summary
Existing single-filter feature selection methods rely on a large amount of labeled fault type DGA data, resulting in poor model generalization performance and reduced fault diagnosis rate. Furthermore, it is difficult to obtain reliable labeled DGA samples in the field.
The confidence criterion is used to select high-confidence samples from unlabeled DGA data to expand the training dataset. A multi-filter semi-supervised feature selection method is used to reduce noise introduction and establish a fault diagnosis model.
It improves the accuracy and robustness of the fault diagnosis model, effectively utilizes limited labeled data and more unlabeled data, and solves the problem of single filter feature selection being dependent on labeled data.
Smart Images
Figure CN115881238B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of model training, in particular to a model training method, a transformer fault diagnosis method and related devices. BACKGROUND
[0002] The power transformer is an important power transmission and transformation equipment in the power grid, which undertakes the task of converting voltage and transmitting current. If the transformer has a serious fault during operation, it will endanger the safe and reliable operation of the power grid.
[0003] Dissolved gas analysis (DGA) is the most commonly used method for power transformer fault diagnosis, which can monitor a variety of gas data in oil without power failure, and the obtained data has good continuity. The proportion and content of the composition of the characteristic gas dissolved in the oil can reflect the fault type and severity. In the prior art, a single filter method is used to select features of transformer DGA data and diagnose fault types.
[0004] However, the present inventors found at least the following technical problems in the process of implementing the technical solutions of the embodiments of the present application:
[0005] Since the single filter feature selection requires a large amount of DGA data labeled with fault types to establish a model, too few labeled data will often lead to low generalization performance of the learner and a decrease in fault diagnosis rate. Due to factors such as the lack of disassembled transformers to determine the fault state and the opaque state of transformer fault data, a large amount of manpower and resources are required to obtain reliable labeled DGA fault samples. SUMMARY
[0006] The present application aims to select samples with high confidence from unlabeled DGA data based on a confidence criterion to expand the model training data set and reduce the introduction of noise, thereby solving the technical problem of excessive reliance on labeled DGA data in the single filter feature selection of the prior art.
[0007] The above application objectives are mainly achieved by the following technical solutions:
[0008] In a first aspect, a model training method is used to train a transformer fault diagnosis model. The training method includes:
[0009] Step 1: Obtain a training sample set;
[0010] Step 2: Divide the sample set into an unlabeled training set, a first labeled training set, and a validation set;
[0011] Step 3: constructing a transformer fault diagnosis model of a first FLDA multi-classifier according to candidate features of the sample set and labels corresponding to the first labeled training set;
[0012] Step 4: calculating a multi-classifier-based discrimination score of samples in the unlabeled training set according to the transformer fault diagnosis model of the first FLDA multi-classifier, and predicting labels of samples in the unlabeled training set and the validation set;
[0013] Step 5: selecting samples in the unlabeled training set with confidence exceeding a preset value to join the first labeled training set to form a second labeled training set;
[0014] Step 6: constructing a transformer fault diagnosis model of a second FLDA multi-classifier according to the second labeled training set and labels corresponding thereto;
[0015] Step 7: calculating a multi-classifier-based discrimination score of samples in the unlabeled training set according to the transformer fault diagnosis model of the second FLDA multi-classifier, and predicting labels of samples in the unlabeled training set and the validation set;
[0016] Step 8: determining whether a training termination condition is met, and if yes, terminating the training; otherwise, returning to Step 5.
[0017] In a second aspect, a model training device is used to train a transformer fault diagnosis model, and the device comprises:
[0018] A sample acquisition module is configured to acquire a training sample set.
[0019] A sample division module is configured to divide the sample set into an unlabeled training set, a first labeled training set and a validation set.
[0020] A first model construction module is configured to construct a transformer fault diagnosis model of a first FLDA multi-classifier according to candidate features of the sample set and labels corresponding to the first labeled training set.
[0021] A first calculation module is configured to calculate a multi-classifier-based discrimination score of samples in the unlabeled training set according to the transformer fault diagnosis model of the first FLDA multi-classifier, and predict labels of samples in the unlabeled training set and the validation set.
[0022] A training set expansion module is configured to select samples in the unlabeled training set with confidence exceeding a preset value to join the first labeled training set to form a second labeled training set.
[0023] A second model construction module is configured to construct a transformer fault diagnosis model of a second FLDA multi-classifier according to the second labeled training set and labels corresponding thereto.
[0024] a second calculation module: calculating the sample's discriminant score based on the multi-classifier in the unlabeled training set according to the transformer fault diagnosis model of the second FLDA multi-classifier, and predicting the labels of the samples in the unlabeled training set and the validation set;
[0025] a judgment module: judging whether the training termination condition is met, if yes, the training is terminated; otherwise, going to the training set expansion module.
[0026] In a third aspect, a transformer fault diagnosis method comprises:
[0027] obtaining DGA data to be diagnosed;
[0028] inputting the DGA data to be diagnosed into a transformer fault diagnosis model to obtain an optimal feature subset corresponding to the DGA data to be diagnosed;
[0029] The transformer fault diagnosis model is trained by using the above model training method.
[0030] In a fourth aspect, a transformer fault diagnosis device comprises:
[0031] a DGA data module for obtaining DGA data to be diagnosed;
[0032] a fault diagnosis module for inputting the DGA data to be diagnosed into a transformer fault diagnosis model to obtain an optimal feature subset corresponding to the DGA data to be diagnosed;
[0033] The transformer fault diagnosis model is trained by using the above model training method.
[0034] Compared with the prior art, the transformer fault diagnosis method adopting the multi-filter semi-supervised feature selection has the beneficial effects that: since the model training adopts the confidence criterion to select samples with high confidence in the unlabeled DGA data to expand the training data set, the introduction of noise is reduced, so that the fault diagnosis model can be established by using limited labeled DGA data and more unlabeled DGA data, thereby solving the technical problem that the single filter feature selection in the prior art excessively depends on the labeled DGA data. BRIEF DESCRIPTION OF DRAWINGS
[0035] Figure 1 A flowchart of a model training method provided by an embodiment of the present application is shown in the figure;
[0036] Figure 2 A flowchart of a model training method provided by an embodiment of the present application is shown in the figure;
[0037] Figure 3 A structural diagram of a model training device provided by an embodiment of the present application is shown in the figure;
[0038] Figure 4 A flowchart of a transformer fault diagnosis method provided for Embodiment Four of the present application is shown in the figure.
[0039] Figure 5 A structural diagram of a transformer fault diagnosis device provided for Embodiment Five of the present application is shown in the figure. DETAILED DESCRIPTION
[0040] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the present application.
[0041] Embodiment One
[0042] A model training method is used to train a transformer fault diagnosis model, as shown in the figure, comprising the following steps: Figure 1
[0043] Step 1: Select DGA data to form a sample set for model training.
[0044] It should be noted that the transformer faults based on DGA are divided into five categories: low-energy discharge (LED), high-energy discharge (HED), low-to-moderate temperature overheating (LMT, below 700℃), high-temperature overheating (HT, above 700℃), and normal operation (N).
[0045] In the embodiments of the present application, the sample feature dimension is selected to be 64 dimensions (as shown in Table 1 below), and the adaptive synthetic sampling algorithm (ADASYN, Adaptive Synthetic Sampling Algorithm) is used to increase the sample number from 158 to 242, balancing the sample distribution. The main idea of the ADASYN algorithm is to use the density distribution as the standard to autonomously determine the number of synthetic samples to be generated for each minority class sample.
[0046] Then, the candidate features of the balanced samples are subjected to arctangent transformation (AT, Arctangent Transformation):
[0047]
[0048] and normalization processing:
[0049]
[0050] where xij and x ij respectively represent the jth feature value of the ith sample before and after normalization, x j max and x j min respectively represent the maximum value and the minimum value of the jth feature of all samples.
[0051]
[0052] Table 1 DGA data candidate feature set
[0053] Step 2: divide the sample set into an unlabeled training set, a first labeled training set and a validation set.
[0054] The above pre-processed DGA sample set is divided into a training set D (194 samples) and a validation set D V (48 samples) according to 20*5 cross-validation, and the training set is divided into a labeled training set D I (75 samples) and an unlabeled extended training set D F (119 samples).
[0055] 20*5 cross-validation means that the sample set is divided into 5 parts on average, and 4 of them are used as training sets in turn, and the remaining 1 is used as an independent test set. The classification set is randomly sorted for 20 times, and 100 different combinations of training sets and validation sets are formed.
[0056] Step 3: According to the candidate features of the sample set and the labels corresponding to the first labeled training set, a transformer fault diagnosis model of the first FLDA multi-classifier is constructed.
[0057] Step 4: Calculate the multi-classifier-based discriminant score of the samples in the unlabeled training set according to the transformer fault diagnosis model of the first FLDA multi-classifier, and predict the labels of the samples in the unlabeled training set and the validation set.
[0058] Using the candidate features shown in Table 1 and the labels corresponding to the labeled training set D I (the first labeled training set), a transformer fault diagnosis model of a Fisher linear discrimination analysis (FLDA) multi-classifier (the transformer fault diagnosis model of the first FLDA multi-classifier) is established based on one-to-one (OAO, One-against-one), the multi-classifier-based discriminant score of the samples in the unlabeled training set D F is calculated, and the labels of the samples in the unlabeled training set D F and the validation set D V are predicted, and the iteration number m=0.
[0059] The goal of FLDA is to find a transformation matrix W that maximizes the distance between classes and minimizes the distance within classes. The discriminant score f(x) is calculated using the transformation matrix W and the parameter b as follows:
[0060]
[0061]
[0062]
[0063]
[0064] f(x) = Wx + b;
[0065] FLDA is used to solve the binary classification problem, while transformer fault classification is a multi-classification problem. Therefore, it is necessary to extend FLDA to multi-classification FLDA. In the embodiment of the present application, OAO is used to realize multi-classification. OAO is used to train n*(n-1) / 2 binary classifiers, where n represents the number of classes. After training, the discriminant score f j (x) is calculated for all unlabeled samples based on the binary classifier, where j = 1, 2,..., n*(n-1) / 2. When classifying the unlabeled samples, the class with the most votes is the classification result of the unlabeled sample.
[0066] Step 5: Select the samples with confidence exceeding the preset value in the unlabeled training set to add to the first labeled training set to form a second labeled training set.
[0067] In the mth (m = 1, 2,...) iteration, the samples with high confidence in the unlabeled training set D F are added to the labeled training set D I to form a new labeled training set Q(m), and the label of Q(m) is denoted as y m (.).
[0068] Confidence evaluation criterion:
[0069] In the semi-supervised learning process, there are two reasons that hinder the performance of the classifier: one is that using a small amount of labeled training set cannot obtain a reliable initial model, and the other is that the amount of information of the unlabeled samples used to update the initial model is insufficient. Therefore, it is crucial to find some high-confidence data from the unlabeled data. The embodiment of the present application proposes a multi-class nearest average distance (MNAD, Multi-class Nearest Average Distance) criterion to find high-confidence data from the unlabeled training set D FThe labeled training set is expanded by selecting high-confidence samples. The criterion is achieved by analyzing the discriminant scores of FLDA multi-classifier on unlabeled samples. In the proposed MNAD criterion, FLDA binary classifiers are trained based on the labeled training set using OAO strategy, and voting is performed according to multiple discriminant scores (j = 1, 2, …, n*(n-1) / 2) of each unlabeled sample, only considering the samples that obtain all votes (at most n-1 votes for each class), and then selecting the samples that satisfy the following formula to expand the labeled training set in each iteration.
[0070] S = Vote (X) full (X ∈ D F );
[0071] value = mean(scores(S)) (S ∈ D S );
[0072] selectsamples = {value > h*mean(value)};
[0073] where S represents the samples that obtain all votes from the unlabeled sample set, value and h represent the average decision score and weight parameter respectively.
[0074] Step 6: Construct a transformer fault diagnosis model of the second FLDA multi-classifier according to the second labeled training set and its corresponding labels.
[0075] Step 7: Calculate the discriminant scores of the samples in the unlabeled training set based on the multi-classifier according to the transformer fault diagnosis model of the second FLDA multi-classifier, and predict the labels of the samples in the unlabeled training set and the validation set.
[0076] Using the new labeled training set Q(m) and its corresponding labels, a transformer fault diagnosis model of the FLDA multi-classifier is established, the discriminant scores of the samples in the unlabeled training set D F are recalculated based on the multi-classifier, and the labels of the samples in the unlabeled training set D F and the validation set D V are re-predicted.
[0077] Step 8: Determine whether the training termination condition is met. If yes, the training is terminated; otherwise, return to step 5.
[0078] When the number of iterations m is greater than 1, the discriminant scores of the samples in the unlabeled training set D FThe number of samples t(m) of different prediction labels, when the iteration satisfies the preset maximum number of iterations or t(m) = 0, the training is terminated, otherwise, go to step 5 to continue the iteration training of the m+1th time. In the embodiment of the application, the maximum number of iterations is set to 10, because the accuracy tends to be stable after 10 iterations.
[0079] t(m) = number(find(y m (.) ≠ y m-1 (.)))(m > 1);
[0080] Wherein, number(.) represents the number of samples satisfying the condition in the bracket.
[0081] Compared with the technical effect obtained by the prior art: the transformer fault diagnosis method adopting the multi-filter semi-supervised feature selection of the application, since the model training adopts the confidence criterion to select the samples with high confidence in the unmarked DGA data to expand the training data set, reduces the introduction of noise, so that the limited marked DGA data and more unmarked DGA data can be used to establish the fault diagnosis model, so as to solve the technical problem that the single filter feature selection in the prior art excessively depends on the marked DGA data, and the unmarked DGA data is easy to obtain on site but cannot be used.
[0082] Embodiment two
[0083] The model training method is used for training the transformer fault diagnosis model, as shown in Figure 2 , comprising the following steps:
[0084] Step 1: selecting DGA data to form a sample set for model training.
[0085] Step 2: dividing the sample set into an unmarked training set, a first marked training set and a validation set.
[0086] Step 3: constructing a transformer fault diagnosis model of a first FLDA multi-classifier according to the candidate features of the sample set and the labels corresponding to the first marked training set.
[0087] Step 4: calculating the multi-classifier-based discrimination score of the samples in the unmarked training set according to the transformer fault diagnosis model of the first FLDA multi-classifier, and predicting the labels of the samples in the unmarked training set and the validation set.
[0088] Step 5: selecting the samples with confidence exceeding a preset value in the unmarked training set to join the first marked training set to form a second marked training set.
[0089] Step 6: constructing a transformer fault diagnosis model of a second FLDA multi-classifier according to the second marked training set and the labels corresponding thereto.
[0090] Step 7: Calculate the multi-classifier-based discriminant scores of the samples in the unlabeled training set according to the transformer fault diagnosis model of the second FLDA multi-classifier, and predict the labels of the samples in the unlabeled training set and the validation set.
[0091] Step 8: Determine whether the training termination condition is met. If yes, terminate the training; otherwise, return to Step 5.
[0092] Steps 1-8 in Embodiment One have been described in detail and will not be repeated here.
[0093] Step 9: Calculate the final iteration accuracy of the samples in the validation set, and use multiple different filters to sort the candidate features of the second labeled training set based on the accuracy.
[0094] After the training is terminated, the final iteration accuracy of the samples in the validation set D V is calculated.
[0095] According to the candidate features of the new training set Q(m) and the corresponding final labels, use multiple different filters: ReliefF, Mutual Information (MI), Infinite Latent Feature Selection (ILFS), Max-Relevance and Min-Redundancy (MRMR), and Least absolute shrinkage and selection operator (Lasso) to sort the candidate features.
[0096] Step 10: Calculate the average accuracy of the samples in the validation set through r*k-fold cross-validation; obtain the r*k times of candidate feature sorting results of the second labeled training set for each filter, form a Rank matrix, and the size of the matrix is (r*k)*c, c represents the number of candidate features of the second labeled training set, (r*k) and c represent the number of rows and columns of the matrix, respectively.
[0097] Step 11: According to the matrix, select the optimal feature subset using the feature combination method based on the proportional wrapping strategy.
[0098] Through 20*5-fold cross-validation, the average accuracy of the samples in the validation set is calculated. For each filter method, a Rank matrix (matrix size is 100*64) of 100 sorting results is obtained. According to the result matrix, the optimal feature combination is selected using the feature combination method based on the proportional wrapping strategy.
[0099] Specifically, for each filter method, the sum of each column of the computed result matrix is calculated, and all features are ranked from small to large. After semi-supervised learning using different proportions of features is performed with 20*5 cross-validation, the feature subset with the best average accuracy on the validation set D V The feature subset with the best average accuracy on the validation set D
[0100] Based on the same initial labeled training set, the average accuracy of transformer fault diagnosis on the validation set after 20*5 cross-validation of the multi-filter standard feature selection, single-filter standard feature selection, standard no-feature selection, multi-filter semi-supervised feature selection, single-filter semi-supervised feature selection, and semi-supervised no-feature selection algorithms of the embodiment of the application is compared. The single-filter selects MI as the filter. As shown in Table 2.
[0101] Algorithm Mean accuracy (%) Selected optimal feature subset number Multi-filter standard feature selection 85.85 57,58,59,60,61,62,63,64 Single-filter standard feature selection 84.12 10,28,35,38,45,48,59,62 Standard no-feature selection 62 1-64 Multi-filter semi-supervised feature selection 88.49 55,56,57,58,59,50,61,62,63,64 Single-filter semi-supervised feature selection 86.23 27,30,35,39,48,49,50,59,61,63 Semi-supervised no-feature selection 55 1-64
[0102] Table 2: Average accuracy of transformer fault diagnosis obtained by different algorithms and optimal feature subset number
[0103] As shown in Table 2, except for the no-feature selection algorithm, whether based on a multi-filter or a single-filter, the semi-supervised method can obtain a higher average accuracy than the standard method, which shows that the semi-supervised algorithm that expands the labeled sample set by using unlabeled samples can optimize the model of the standard algorithm and improve the fault diagnosis rate. Whether for the standard algorithm or the semi-supervised algorithm, the algorithm using the multi-filter feature selection can better evaluate the correlation between the features and the categories than the single-filter feature selection, obtain a better optimal feature subset, and further improve the fault diagnosis rate. However, for the no-feature selection algorithm, the semi-supervised learning obtains a lower average accuracy than the standard algorithm, because the redundancy of the features leads to the weakness of the semi-supervised initial model, and with the increase of iterations, the noise introduced will accumulate, resulting in a decrease in the accuracy, which also shows the importance of feature selection in semi-supervised learning.
[0104] The DGA data set provided by the local power company is used to test the robustness of the optimal feature subset selected by the proposed method, as shown in Table 3:
[0105]
[0106]
[0107] Table 3: Features selected by different methods and corresponding semi-supervised average accuracy
[0108] Table 3 shows the feature subsets selected by different traditional methods and the proposed method, and the average fault diagnosis rates obtained after semi-supervised 20*5 cross-validation on the DGA dataset provided by the power company. Obviously, the fault diagnosis method of the multi-filter semi-supervised feature selection obtains a much higher average accuracy than other traditional methods, indicating the strong robustness of the optimal feature subset selected by the proposed algorithm.
[0109] As can be seen from the above, the embodiments of the present application test the method on the IEC T10 public dataset and compare it with the traditional supervised diagnosis model. The results show that the method can better optimize the DGA features and significantly improve the accuracy of transformer fault diagnosis. In addition, by testing the DGA sample set of the local power supply company, the strong robustness of the optimal feature subset is verified.
[0110] Since the single filter feature selection method in the prior art cannot comprehensively and fairly use all information to evaluate the correlation between features and categories, it may even lead to a suboptimal feature subset. Therefore, the transformer fault diagnosis method based on multi-filter semi-supervised feature selection in the embodiments of the present application adopts five filtering technologies to sort the candidate DGA features under the framework of semi-supervised learning, and then uses a feature combination method based on the proportional wrapping strategy to fuse the features of multiple filters to form an optimal feature subset. The problem of the single filter feature selection method in the prior art that cannot comprehensively and fairly use all information to evaluate the correlation between features and categories, which may even lead to a suboptimal feature subset, is solved, and the problem of data imbalance is improved.
[0111] Embodiment three
[0112] The model training device is used for training a transformer fault diagnosis model, as shown in Figure 3 The device comprises:
[0113] The sample acquisition module is used for acquiring a training sample set.
[0114] The sample division module is used for dividing the sample set into an unmarked training set, a first labeled training set and a validation set.
[0115] The first model construction module is used for constructing a transformer fault diagnosis model of a first FLDA multi-classifier according to the candidate features of the sample set and the labels corresponding to the first labeled training set.
[0116] The first calculation module is used for calculating the multi-classifier-based discrimination scores of the samples in the unmarked training set according to the transformer fault diagnosis model of the first FLDA multi-classifier, and predicting the labels of the samples in the unmarked training set and the validation set.
[0117] The training set expansion module selects samples with confidence exceeding a preset value in the unlabeled training set to join the first labeled training set to form a second labeled training set.
[0118] The second model construction module constructs a transformer fault diagnosis model of the second FLDA multi-classifier according to the second labeled training set and the corresponding labels.
[0119] The second calculation module calculates the discriminant scores of the samples in the unlabeled training set based on the multi-classifier according to the transformer fault diagnosis model of the second FLDA multi-classifier, and predicts the labels of the samples in the unlabeled training set and the validation set.
[0120] The judgment module judges whether the training termination condition is met, and if yes, the training is terminated; otherwise, the training set expansion module is turned to.
[0121] In a preferred embodiment, the judgment module specifically judges the rule as follows: when any one of the following two conditions is met, it is determined that the training termination condition is met, otherwise it is determined that the training termination condition is not met: the iteration reaches a preset maximum iteration number or t(m) = 0; wherein t(m) is the number of samples with predicted labels in the unlabeled training set in any adjacent two consecutive iterations.
[0122] In a preferred embodiment, the device further comprises a third calculation module: calculating the final iteration accuracy rate in the validation set, and sorting the candidate features of the second labeled training set based on the accuracy rate using multiple different filters.
[0123] In a preferred embodiment, the device further comprises a fourth calculation module and an optimal feature subset selection module, wherein the fourth calculation module is used to: calculate the average accuracy rate of the samples in the validation set through r*k-fold cross-validation; obtain r*k times of the candidate feature sorting results of the second labeled training set for each filter to form a Rank matrix, the size of the matrix is (r*k)*c, c represents the number of candidate features of the second labeled training set, (r*k) and c represent the number of rows and columns of the matrix respectively; and the optimal feature subset selection module is used to: select the optimal feature subset according to the matrix by using a proportional wrapping strategy feature combination method.
[0124] The model training device of the embodiment of the present application corresponds to the model training method of the above-mentioned embodiment, and realizes the corresponding functions. Since the steps of the model training method have been described in detail in the above-mentioned embodiment, they will not be repeated here.
[0125] Embodiment Four
[0126] The transformer fault diagnosis method, for example,Figure 4 As shown, comprising the following steps:
[0127] Obtaining DGA data to be diagnosed;
[0128] Inputting the DGA data to be diagnosed into a transformer fault diagnosis model to obtain an optimal feature subset corresponding to the DGA data to be diagnosed;
[0129] The transformer fault diagnosis model is obtained by training using the model training method.
[0130] The specific training method of the transformer fault diagnosis model can refer to embodiments one and two.
[0131] Embodiment five
[0132] A transformer fault diagnosis device, such as Figure 5 As shown, comprising:
[0133] A DGA data module for obtaining DGA data to be diagnosed;
[0134] A fault diagnosis module for inputting the DGA data to be diagnosed into a transformer fault diagnosis model to obtain an optimal feature subset corresponding to the DGA data to be diagnosed;
[0135] The transformer fault diagnosis model is obtained by training using the model training method.
[0136] The specific training method of the transformer fault diagnosis model can refer to embodiments one and two.
[0137] It should be noted that for the above-mentioned method embodiments, in order to simply describe, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited by the described action sequence, because according to the present application, certain steps can be performed in other order or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily necessary for the present application.
[0138] In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can refer to the related description of other embodiments.
[0139] In several embodiments provided in the present application, it should be understood that the disclosed apparatus can be implemented in other manners. For example, the division of the apparatus embodiments described above is merely illustrative, and the division of the units can be different, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.
[0140] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.
[0141] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can be physically present separately, or two or more units can be integrated into one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.
[0142] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable memory. Based on this understanding, the technical solutions of the present application essentially or the part of the prior art that contributes to the technical solutions or all or part of the technical solutions can be embodied in the form of a software product, which is stored in a memory and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned memory includes: a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.
[0143] A person of ordinary skill in the art can understand that all or part of the steps of the various methods of the above embodiments can be completed by a program instructing relevant hardware, and the program can be stored in a computer readable memory, which can include a flash disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.
[0144] The above has described the embodiments of the present application in detail, and the principles and implementation manners of the present application are described by applying specific examples. The above description of the embodiments is only used to help understand the method of the present application and its core idea. Meanwhile, for the general technical personnel in the art, according to the idea of the present application, the specific implementation manners and application ranges will be changed, and the above description should not be understood as the limitation of the present application.
Claims
1. A model training method characterized by, The application discloses a transformer fault diagnosis model training method and device. Step 1: obtaining a training sample set; Step 2: dividing the sample set into an unlabeled training set, a first labeled training set and a validation set; Step 3: constructing a transformer fault diagnosis model of a first FLDA multi-classifier according to candidate features of the sample set and labels corresponding to the first labeled training set; Step 4: calculating a multi-classifier-based discrimination score of samples in the unlabeled training set based on the transformer fault diagnosis model of the first FLDA multi-classifier, and predicting labels of samples in the unlabeled training set and the validation set; Step 5: selecting samples with confidence exceeding a preset value in the unlabeled training set to join the first labeled training set to form a second labeled training set; Step 6: constructing a transformer fault diagnosis model of a second FLDA multi-classifier according to the second labeled training set and labels corresponding thereto; Step 7: calculating a multi-classifier-based discrimination score of samples in the unlabeled training set based on the transformer fault diagnosis model of the second FLDA multi-classifier, and predicting labels of samples in the unlabeled training set and the validation set; Step 8: determining whether a training termination condition is met, and if yes, terminating the training; otherwise, returning to Step 5; Step 9: calculating a final iteration accuracy of the validation set, and sorting candidate features of the second labeled training set based on the accuracy by using multiple different filters; The multiple different filters include a ReliefF, a mutual information, an unlimited potential feature selection, a maximum correlation minimum redundancy and a minimum absolute value convergence and selection operator; Step 10: Calculate the average accuracy of the samples in the validation set through 20*5-fold cross-validation; obtain the candidate feature ranking results of the second labeled training set for 20*5 times for each filter to form a Rank matrix, the size of the matrix is ( 20 * 5 * c , c , where N represents the number of candidate features of the second labeled training set, 20 * 5 , and c represent the number of rows and columns of the matrix, respectively; Step 11: five different filter methods obtain five optimal feature subsets; optimal feature subsets of different filter methods are voted, features with a frequency of being selected more than twice are retained to form a new optimal feature subset, and otherwise discarded. 2.The model training method of claim 1, wherein, The determination whether the training termination condition is met specifically includes: When any one of the following two conditions is satisfied, it is determined that the training termination condition is satisfied, otherwise it is determined that the training termination condition is not satisfied: the iteration reaches a preset maximum iteration number or t ( m ) = 0; wherein t ( m ) is the number of samples of the predicted label in the unmarked training set of any two adjacent continuous iterations.
3. The model training apparatus, characterized by, The application discloses a transformer fault diagnosis model training device. A sample obtaining module is used to obtain a training sample set; A sample dividing module is used to divide the sample set into an unlabeled training set, a first labeled training set and a validation set; A first model constructing module is used to construct a transformer fault diagnosis model of a first FLDA multi-classifier according to candidate features of the sample set and labels corresponding to the first labeled training set; A first calculating module is used to calculate a multi-classifier-based discrimination score of samples in the unlabeled training set based on the transformer fault diagnosis model of the first FLDA multi-classifier, and predict labels of samples in the unlabeled training set and the validation set; A training set expanding module is used to select samples with confidence exceeding a preset value in the unlabeled training set to join the first labeled training set to form a second labeled training set; A second model constructing module is used to construct a transformer fault diagnosis model of a second FLDA multi-classifier according to the second labeled training set and labels corresponding thereto; A second calculation module: calculating the discriminant score of the samples in the unlabeled training set based on the multi-classifier according to the transformer fault diagnosis model of the second FLDA multi-classifier, and predicting the labels of the samples in the unlabeled training set and the validation set; A judgment module: judging whether the training termination condition is met, if yes, the training is terminated; otherwise, going to the training set expansion module; A third calculation module: calculating the final iteration accuracy of the validation set, and sorting the candidate features of the second labeled training set using a plurality of different filters based on the accuracy; The plurality of different filters include: ReliefF, mutual information, infinite potential feature selection, maximum correlation minimum redundancy, and minimum absolute value convergence and selection operator; The fourth calculation module: After 20*5 fold cross-validation, calculates the average accuracy of the samples in the validation set; for each filter, obtains the candidate feature ranking results of the second labeled training set 20*5 times, forming a Rank matrix, the size of which is ( 20 * 5 )* c , c This represents the number of candidate features in the second labeled training set. 20 * 5 )and c These represent the number of rows and columns of the matrix, respectively. An optimal feature subset selection module: five different filter methods obtain five optimal feature subsets; the optimal feature subsets of different filter methods are voted, the features selected more than twice are retained to form a new optimal feature subset, otherwise discarded. 4.The model training apparatus of claim 3, wherein The specific judgment rule of the judgment module is: when any one of the following two conditions is met, it is determined that the training termination condition is met, otherwise it is determined that the training termination condition is not met: the iteration reaches a preset maximum iteration number or t ( m )=0; wherein t ( m ) is the number of samples of the predicted label in the unmarked training set of any two adjacent continuous iterations.
5. A method of diagnosing a fault in a transformer, characterized by, It comprises: Obtaining DGA data to be diagnosed; Inputting the DGA data to be diagnosed into the transformer fault diagnosis model to obtain the optimal feature subset corresponding to the DGA data to be diagnosed; Wherein the transformer fault diagnosis model is trained by the model training method of any one of claims 1-2.
6. A transformer fault diagnostic device characterized by, It comprises: A DGA data module for obtaining DGA data to be diagnosed; A fault diagnosis module for inputting the DGA data to be diagnosed into the transformer fault diagnosis model to obtain the optimal feature subset corresponding to the DGA data to be diagnosed; Wherein the transformer fault diagnosis model is trained by the model training method of any one of claims 1-2.